Microsoft Research has released Orchard, an open-source framework that trains AI agents to solve complex tasks like coding and web navigation using remarkably small models. The breakthrough centers on a Kubernetes-based environment that lets developers train agents directly inside real deployment systems, eliminating the gap between training and production that has plagued agent development. Orchard-SWE, the framework's software engineering variant, achieves 73% accuracy on the SWE-bench Verified benchmark using only 3 billion active parameters, approaching the performance of systems using more than 10 times larger models.

What Makes Microsoft Orchard Different From Other Agent Frameworks?

Most AI agent frameworks focus on orchestrating and deploying agents after they're built. Orchard takes a different approach by providing the infrastructure to train agents from scratch with production-ready quality. The framework's core component, Orchard Env, is a lightweight Kubernetes service that creates, manages, and removes thousands of isolated agent environments in parallel. This enables researchers and engineers to run reinforcement learning experiments at scale without building custom infrastructure for each task type.

The framework supports three specialized agent workflows: Orchard-SWE for software engineering tasks like bug fixing and code generation, Orchard-GUI for web navigation, and Orchard-Claw for personal assistance. Each workflow demonstrates that smaller open-weight models can achieve strong results on complex real-world tasks. What sets Orchard apart is its ability to train agents directly inside real deployment harnesses such as Claude Code, Codex, OpenClaw, and ZeroClaw, ensuring that capabilities learned during training transfer seamlessly to production environments.

How Does Orchard Achieve Such Strong Results With Smaller Models?

Orchard's success comes from three advanced training techniques that work together to maximize learning efficiency. The framework uses curriculum learning, which dynamically adjusts the number of practice attempts based on task difficulty. Easier problems receive fewer rollouts, while challenging tasks get more opportunities for the agent to discover successful solutions. This prevents wasting computational resources on problems the agent has already mastered.

Instead of sparse binary rewards that only tell an agent whether it succeeded or failed, Orchard-SWE uses dense reward signals that provide feedback at each step of the problem-solving process. This approach accelerates learning and produces more reliable agent behavior. Additionally, a compact 4-billion-parameter value model trained on past rollouts recognizes high-quality solutions. At problem-solving time, this model scores several candidate answers and selects the best one, boosting performance from 69.7% to 73% on SWE-bench Verified.

The results across multiple domains show the framework's versatility. Orchard-GUI achieves 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, for an average of 68.4%, using only 400 distilled demonstrations combined with 2,200 open-ended training tasks. Orchard-Claw reaches 59.6% success rate on Claw-Eval with up to three attempts, increasing to 73.9% when paired with the ZeroClaw agent system. Training across multiple harnesses substantially improves reliability.

Steps to Get Started With Microsoft Orchard

• Access the Repository: Download Orchard from GitHub, which includes Orchard Env, training workflows for SWE, GUI, and Claw agents, along with complete training data and evaluation methods for transparency and reproducibility.
• Deploy the Environment Service: Set up Orchard Env on a Kubernetes cluster to handle isolation, scaling, and lifecycle management for agent rollouts, or use lightweight containers for rapid local development and iteration.
• Select Your Workflow: Choose Orchard-SWE for code generation and bug fixing, Orchard-GUI for web navigation tasks, or Orchard-Claw for personal assistance workflows based on your use case.
• Run the Training Pipeline: Execute the training pipeline with your chosen configuration, supporting reinforcement learning, supervised fine-tuning, and value model training with automatic evaluation metrics.

Microsoft provides practical guidance for teams implementing Orchard. Start with Orchard-GUI for faster iteration since web tasks require less compute than code generation. Use the provided training data as a baseline before collecting custom demonstrations specific to your domain. Enable value model reranking for production deployments where accuracy matters most. Monitor resource usage during large-scale rollouts using Orchard Env's built-in observability tools.

The release of Orchard with complete training data and evaluation methods represents a significant shift toward transparency in agent research. Microsoft is enabling the broader research community to build and study open agentic systems using the same tools and data that powered Orchard's results. This open-source approach contrasts with proprietary agent frameworks and positions Orchard as a foundation for future agent development across academia and industry.

For organizations building AI agents, Orchard's parameter efficiency has immediate practical implications. Smaller models cost less to train and run, making agent development more accessible to teams without massive computational budgets. The framework's focus on training infrastructure rather than deployment orchestration fills a gap in the current agent ecosystem, where most tools focus on composing and deploying agents rather than training them from scratch with production-ready quality.